Granite Code
IBM's open code models from 3B to 34B parameters for coding tasks
Granite Code is a family of decoder-only language models built by IBM for software work, covering code generation, bug fixing, explanation and repository tasks. Announced in May 2024 and trained on code in 116 programming languages, the models run from 3 to 34 billion parameters and were released under Apache 2.0 for research and commercial use.
- Sizes range from 3 to 34 billion parameters
- Pretrained on code in 116 programming languages
- Released under Apache 2.0 for research and commercial use
- Trained on IBM's Vela and Blue Vela GPU clusters
- Marked deprecated by IBM in favour of newer Granite models
What it is
Granite Code is a set of open language models from IBM aimed at programming work. The family spans 3 billion to 34 billion parameters, so it covers both large application modernisation jobs and small on-device uses where memory is tight. The models handle code generation, bug fixing, code explanation and documentation, and repository maintenance. IBM says the models were tuned for enterprise software development workflows. The largest member, Granite-34B-Code-Instruct-8K, was released on 6 May 2024 and is a fine-tuned version of Granite-34B-Code-Base. All models in the family were published under the Apache 2.0 licence for both research and commercial use.
How it came about
The family was presented in a paper by a large IBM author team, with Mayank Mishra, Matt Stallone and Gaoyuan Zhang as equal contributors and Rameswar Panda and Ruchir Puri as corresponding authors. The paper was posted to arXiv on 7 May 2024. The 34B instruct model card gives a release date of 6 May 2024. IBM has since marked these models as deprecated and points users to newer mainline Granite releases.
Inside the model
Granite Code models are decoder-only transformers, meaning they predict text forward from what came before. The base models were pretrained on code written in 116 programming languages. The instruct versions were then fine-tuned on permissively licensed instruction data. For the 34B instruct model this mix included code commit data from CommitPackFT covering 92 languages, math sets such as MathInstruct and MetaMathQA, code instruction sets including Glaive-Code-Assistant-v3 and function calling data, and general instruction sets such as HelpSteer and a licence-filtered Platypus. Some math subsets were dropped over licence concerns. Training ran on IBM's Vela and Blue Vela supercomputing clusters using NVIDIA A100 and H100 GPUs across thousands of cards.
What it can do
IBM reports that the family performs consistently well against other open-source code models across tasks such as generation, fixing and explanation, and describes it as an all-round code model. The instruct models are built to follow coding instructions and support building coding assistants, with tuning aimed at logical reasoning and problem solving. Performance may be weaker on programming languages outside the fine-tuning set, and few-shot examples can help steer output there. IBM advises safety testing and task-specific tuning before use in critical applications.
How to use it
All Granite Code models are downloadable under the Apache 2.0 licence, allowing research and commercial use. The 34B instruct model is available on Hugging Face as ibm-granite/granite-34b-code-instruct-8k, with a GitHub repository at ibm-granite/granite-code-models. IBM now labels these models deprecated and recommends the latest mainline Granite family for new projects.
Safety and testing
IBM's model card asks developers to run their own safety testing and target-specific tuning before deploying the models on critical applications. It notes that instruct models inherit ethical considerations and limitations from the base model. No specific red-teaming results or incidents are described.
Reception and debate
IBM positions the family as competitive among open-source code models on a broad set of tasks. The company has since deprecated the line in favour of newer Granite releases, making it available mainly for historical and scientific purposes.
Why it matters
The family was IBM's open offering for code intelligence, sized to cover both server and memory-limited settings. IBM now directs users to its mainline Granite language models, whose code capabilities it says supersede these models.
Questions people ask about Granite Code
What is Granite Code for?
It is a family of code-focused language models from IBM meant for software work: generating code, fixing bugs, explaining and documenting code, and maintaining repositories. IBM says the models were tuned for enterprise development workflows and can back coding assistants.
How big are the models?
The family runs from 3 billion to 34 billion parameters. Smaller sizes target on-device or memory-limited use, while the largest, 34B, is aimed at heavier tasks such as application modernisation.
Can I use it commercially?
Yes. IBM released all Granite Code models under the Apache 2.0 licence, which covers both research and commercial use, and the weights can be downloaded from Hugging Face.
Is it still recommended?
No. IBM marks these models deprecated and not recommended for new projects, pointing users to the latest mainline Granite language model family, whose code abilities it says supersede this line. The old models stay available for historical and scientific use.
What data was it trained on?
Base models saw code in 116 programming languages. Instruct versions added permissively licensed instruction data: code commits from CommitPackFT across 92 languages, math sets, code and function-calling instructions, and general instruction sets such as HelpSteer and Platypus.
Good reading and sources
There is no English Wikipedia article on Granite Code yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2405.04324, huggingface.co/ibm-granite/granite-34b-code-instruct-8k. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.